Forecasting: AI plus hygiene

The sales world is having an AI moment. Forecasting is the part everyone wants fixed first. It is easy to see why. A forecast is not just a spreadsheet number. It is the number that drives hiring plans, board conversations, and whether a company thinks it can spend. That is why the current push for AI-assisted forecasting comes with a less glamorous demand right beside it: cleaner pipeline data. The old version of forecast math is simple. Take every open deal, assign a probability based on stage, and roll it up into a weighted pipeline. If a $1,000 deal sits in a 50% stage, the forecast counts $500, not $1,000. Sales tools have formalized that logic for years. Salesloft describes weighted pipeline exactly that way, as the sum of deal amounts adjusted by stage probability. (help.salesloft.com) But that neat arithmetic breaks fast when stage labels stop meaning anything. That is the real story behind the new AI push. Modern vendors are promising models that can ingest far more than stage and amount. Salesforce now pitches AI forecasting as a way to combine large datasets and spot patterns humans miss, while still framing forecasting as both data and judgment. It also reduces the problem to three basics: accurate close dates, real customer buying readiness, and probability. (salesforce.com) Those inputs sound obvious. In practice, they are where forecasts go bad. The hygiene problem is not subtle. HubSpot’s forecasting guide says management needs strict pipeline stage definitions, milestones, and data-entry standards because age and stage are often missed, which leads directly to inaccurate forecasts. (hubspot.com) Clari makes the same point in plainer terms: a solid forecast starts with clear definitions for leads, opportunities, and pipeline stages, and those definitions have to be shared across the sales team. (clari.com) If one rep moves a deal to “proposal” after a real commercial review and another does it after sending a first deck, the CRM is no longer a system of record. It is a pile of private interpretations. That is why practitioners are talking about “real” pipeline instead of just large pipeline. The card’s 70–80% target is really a warning against padded opportunity lists. Clari’s own customer community treats idle deals as likely to slip into future quarters or as fluff added to fill out pipeline. (community.clari.com) Gong’s forecasting advice pushes managers to look at overall deal health and use deal intelligence during pipeline reviews, not just accept the headline number. (help.gong.io) AI can help here by scanning for slippage patterns, stale next steps, or close dates that keep moving. But those signals only matter if the underlying fields are updated consistently. That tension explains why a forecasting tournament matters. Metaculus’s Spring 2026 AI Forecasting Benchmark is a bot-only competition built to test how well AI systems predict real-world outcomes. The current season runs through May 6, 2026, with a $50,000 prize pool. (metaculus.com) The point is not sales forecasting specifically. It is to force different models and scaffolds into the open and see which ones actually predict better. The early lesson from those tournaments is almost rude in its simplicity. In Metaculus’s published Q2 2025 results, top human pro forecasters still beat the bot teams, and the biggest driver of bot performance was the underlying model more than elaborate scaffolding. Their best-performing simple bot did one web search, ran a prompt several times, and aggregated the answers. (metaculus.com) That maps neatly onto the sales case. Fancy forecasting layers are useful, but only after the organization decides what a stage means, what milestone proves it, and which deals are real enough to deserve a probability at all.